Key Notes
- Runway has publicly released Praxis-1’s weights and supporting code under Apache 2.0, extending access beyond its early robotics partners.
- The model combines predicted video with robot actions, while released checkpoints target specific single-arm and dual-arm setups.
- Company benchmark results show progress, but robot adaptation, substantial computing requirements and difficulties with ambiguous instructions remain practical limits.
Runway has publicly released Praxis-1, a robot AI model that predicts both future video and the actions a machine should take. The November 11 release makes its weights and supporting code available under Apache 2.0, including for commercial use.
The launch follows a September preview and testing with Noble Machines, Standard Bots and Ultra. It turns an early-partner program into a downloadable model that outside robotics teams can inspect and adapt.
Video Prediction Becomes Robot Control
Praxis-1 uses a 27-billion-parameter mixture-of-experts video backbone, with roughly three billion parameters active per token. An action component reads the video model’s internal features to generate movements alongside its prediction of the scene.
Runway’s research report describes a loop that takes an instruction, camera frames and the robot’s pose, then returns a short sequence of video and actions. The machine executes that sequence and supplies fresh observations for the next prediction.
The approach connects anticipating an outcome with producing the motion that causes it. A convincing imagined scene alone is insufficient: the robot also needs commands that its own hardware can execute accurately.
Released Checkpoints Target Specific Robot Setups
The public code repository lists a DROID checkpoint for a single arm with an external camera and wrist camera, plus a bimanual checkpoint for two YAM arms. A base checkpoint provides a starting point for adapting the model to another rig; it does not run a new robot by itself.
Hugging Face’s LeRobot documentation describes the corresponding integration. Its policy predicts approximately 1.6 seconds of video alongside 24 movement commands at 15 Hz. That gives developers a concrete control interface, rather than treating the release as a general-purpose conversational assistant.
Benchmark Scores Need Their Test Context
Runway reports 46.2% success across 1,200 episodes on NVIDIA’s RoboLab-120 benchmark, compared with 42.9% for FLUX 3 Action and 28.0% for π0.5. Those are company-reported benchmark comparisons, rather than evidence of the same success rate across customer factories or homes.
The research report also identifies failures involving similar-looking objects, vague instructions and long tasks issued through a single prompt. Breaking a project into separate instructions remains different from asking a robot to independently complete a whole workflow.
Open Access Still Requires Adaptation and Compute
The repository’s default setup calls for a CUDA GPU with at least 80 GB of memory. It also describes an experimental lower-precision route for selected workstation GPUs, while warning that its speed and closed-loop accuracy still need measurement. The published hardware requirements deserve more attention than a broad claim that a model can run on many NVIDIA devices.
Our coverage of NVIDIA’s GR00T 2 follows another effort to bring human video into robot learning. Hedra’s Model Factory addresses the training and adaptation work around such systems. Praxis-1 gives developers another model to evaluate, with practical value depending on performance on their own cameras, machines and tasks.
Disclaimer: RobotsBeat is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, Robotics, technology, software, and digital innovation sectors. These relationships do not influence RobotsBeat's editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
